Sabich v2.3: a fifth more recordings, and how it scored
Sabich v2.3 went out on 2026-10-30 and is now the model the Hebrish apps download. It is a Sabich-only update to Release v2: same recipe as v2.2, more recordings. This post covers the training data, how it was scored, and where the data and evaluation files are.
What changed from v2.2 to v2.3
Nothing in the recipe. Like v2.2, v2.3 is retrained from stock Parakeet TDT 0.6B v2 (not on top of v2.2) for 20 epochs on one L40S GPU, with speed perturbation at 0.9x, 1.0x and 1.1x, learning rate 3e-5, and the training mix that includes everything. The run took 91.1 minutes. What changed was the data.
- More recordings: 1,074.3 minutes of the speaker's takes, up from 899.7 for v2.2, about a fifth more. Of those, 522.8 minutes are English with Hebrew words in it (v2.2: 471.4).
- More task data: the vocabulary and tech packs recorded since v2.2.
- Priority words: 680 clips with words the speaker's dictation got wrong, each repeated three times (102.6 extra minutes per epoch).
- LibriSpeech replay: 119.4 minutes of cased, punctuated LibriSpeech, so ordinary English isn't forgotten.
In all, 1,296.3 minutes and 12,405 clips. One in 16 of the speaker's training takes was set aside for validation, used only to pick the checkpoint. Averaging the last five epochs scored 4.76% on validation, against 4.72% for the best single epoch (epoch 18), so the best epoch was kept. The lexicon page charts the mix and how often each word appears in training.
The results
v2.3 was scored at release on the held-out test splits of the same dataset revision, against v2.2 on the same clips. These clips were never trained on, and the code-switching split is term-disjoint: its Hebrew words don't appear anywhere in training. Greedy decoding, normalised WER (lowercase, no punctuation).
| Measure | Clips | Sabich v2.2 | Sabich v2.3 |
|---|---|---|---|
| Code-switching WER | 672 | 5.12% | 4.74% |
| Hebrew words not seen in training, written right | 524 occurrences | 56.3% | 58.0% |
| Own English WER | 571 | 3.90% | 3.48% |
| LibriSpeech WER | 200 | 1.47% | 1.57% |
| Long-form WER | 25 | 6.13% | 6.01% |
| Numbers written right | 84 | 56.0% | 60.7% |
v2.3 is ahead on everything except LibriSpeech. In counts: 304 of 524 unseen Hebrew words right instead of 295, and 51 of 84 numbers instead of 47. In the English set, tech dictation went from 3.58% to 3.26% WER over 348 clips. Named terms in the English set (tech, product and people's names) went from 538 of 654 written right to 553.
What is within noise
These are small steps on one speaker's recordings, so read them as a direction, not a size. LibriSpeech moved 0.1 points on 200 clips, which is within noise. The term gains are spread thinly: most terms gained or lost a single hit, and a few went backwards (kneidlach lost two of six). Some words are still missed by both versions every time: tachles (0 of 7) and freier (0 of 6) both come out in a different spelling. Words like these are what the priority-word set is for.
The data and the evaluation
- Training data: Read-Aloud-Voice-Sentences at revision
ca08e06b, plus LibriSpeech-PC. Access is by request under the terms of use. Every dataset is listed on the datasets page. - Evaluation: the result files, with every reference and transcript, are in the model repository:
evals/r13-all.json(v2.3) andevals/r13-r12-all.json(v2.2 on the same clips). The training config and result are intraining/r13-all/. The model is free for non-commercial use; access is by request on Hugging Face. - Public benchmark: v2.3 hasn't been run on test set v2 yet. Those results will go on the benchmarks page. Until then, the numbers above don't compare with the test set v2 numbers there.
Try it
The apps offer v2.3 through their normal model update and delete v2.2 once it is installed, so no app update is needed. To hear Sabich without installing anything, try the Sabich demo in your browser. The full release note is in the changelog.
